Image Segmentation And Classification Of Skin Lesions Based On Deep Learning

Malignant skin lesions have a high mortality rate,and due to the small difference between the early skin lesions,it is easy to confuse,resulting in delayed treatment.At present,the main method of image segmentation and classification recognition of skin lesions still relies on the experience of professional doctors.According to the color,texture,size,shape and other characteristics of the lesion area,it is more subjective to judge,which is more dependent on professional doctors.Due to the diversity of images,the fatigue and tolerance of doctors are constantly consumed,which makes the accuracy of image segmentation and recognition not guaranteed.Traditional medical image segmentation and classification methods often design feature extraction methods for specific images,resulting in poor generalization ability of the final model,which can not be well applied to other medical images.In order to solve the problems such as fuzzy boundary of skin lesions and imbalance of samples among categories,this paper studies two aspects of disease region segmentation and multi category recognition based on deep learning and lightweight model.The main research work of this paper is as follows:(1)Aiming at the problem of blurred boundaries and large background interference during the segmentation of the skin disease image,this paper proposes an improved U-Net segmentation network based on octave convolution.Based on the U-Net framework,combined with octave convolution,a high and low frequency dual channel is constructed to increase the receptive field and reduce the amount of calculation..In addition,a hierarchical residual skip structure is proposed to compensate for the loss of spatial information and semantic differences caused by the encoding and decoding process,and then an improved U-net segmentation network based on octave convolution is proposed.Experimental results show that,compared with u-net network,the segmentation accuracy of this network is improved by 5.5%on ISIC-2018 Task 1 dataset.At the same time,the improved network is tested on the special PCOS data set of the research group,and the segmentation accuracy is also improved,which proves the applicability of the improved network.(2)In the multi-class recognition of skin disease images,this paper proposes a separable volume integral model based on multi-level feature fusion.First,a deep separable convolutional network with hyperparameters is designed.The separable convolution is used to reduce model parameters and increase the calculation speed.Then,using the idea of multi-level feature fusion,the features of each convolutional layer are spliced and fused to improve the network Feature utilization.For the problem of highly unbalanced distribution of samples in each category of data,data enhancement is used to expand the small sample data to alleviate the tendency of network.In order to get a better network structure,different combinations of hyper parameters are tested.Experiments show that the classification network with a 4-layer structure has the best performance and is better than most methods,and the improved algorithm based on multi-level feature fusion improves the accuracy of 0.98%compared with the original network.